Examining how behaviors flagged as unsafe look different through a welfare lens, and what happens when the question can't be resolved.